Content
73%Weight 40%Scale 1-5Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.
The body is a well-sequenced, validation-heavy workflow with concrete CLI commands, decision tables, and strong error-recovery and exit gating — workflow clarity is excellent. Its weaknesses are repetition of recovery/anti-redundancy caveats across sections and oversized inline evidence-policy paragraphs that belong in the existing reference-file structure. Actionability is high but diluted slightly by abstract risk-based directives.
Suggestions
Deduplicate the 'don't re-run select/check when state is already known' guidance, which is repeated in Step 0, Step 1, Step 2, and the auto-transition section — state it once in the entry-check step and reference it elsewhere.
Move the long evidence-recording / evidence-invalidation policy text (the exit-conditions paragraphs on check evidence, input surfaces, document edits, and `record-check` shadowing) into a reference file like `comet-classic/reference/check-evidence.md`, keeping only the recording command and the guard invocation inline.
Replace vague risk-based directives like "按风险执行相关检查" with a concrete risk-tier table (examples of low/medium/high-risk changes and the corresponding minimum checks), or point to where that grading is defined in `subagent-dispatch.md`.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dense and operational with no beginner-concept padding, but redundancy is more than minor: the anti-redundancy instruction ("不重复 select/check" / don't re-run entry checks) is restated in Steps 0, 1, 2, and the auto-transition section, and "任务未勾选不等于尚未实现" appears twice (lines 30 and 67). The ~500-word evidence-policy paragraph (line 186) and the adjacent evidence/recording paragraphs are oversized and could be tightened or split into a reference file. This fits anchor 3 ('mostly efficient but could be tightened') better than 2 (no padded explanations of known concepts) or 4 (repetition exceeds 'minor'). | 3 / 5 |
Actionability | Concrete executable commands are given throughout — e.g., `comet state set <name> build_mode autonomous subagent_dispatch null tdd_mode tdd review_mode standard --json`, `comet guard <change-name> build --apply`, `comet check run <change-name> build --local -- <program> [args...]` — plus a fully-specified common-case example, config option tables, and a concrete plan-header YAML. However, several directives remain abstract ("按风险执行相关检查", "按风险分级" deferring entirely to external files), keeping it at anchor 4 rather than the fully copy-paste-ready anchor 5. | 4 / 5 |
Workflow Clarity | The workflow is explicitly sequenced (Steps 0 → 1 → 2 → 3 → 3b → 4 → 5 → exit conditions) with pervasive validation checkpoints: entry verification via `comet state check <name> build --json`, BLOCKED/branch-mismatch handling that pauses for a user decision, RED/GREEN evidence requirements with real results, mandatory resolution of CRITICAL/IMPORTANT findings, a complete exit checklist (all tasks checked, code committed, build/tests passed, config fields set), and the `guard --apply` gate with evidence-invalidation feedback loops ("按清单处理，不要凭猜测全量重跑"). This matches anchor 5: clear sequence, explicit validation, and error-recovery loops. | 5 / 5 |
Progressive Disclosure | The skill ships no bundle files, but the body clearly signals one-level-deep external references ("按 `comet-classic/reference/context-recovery.md` 执行", `debug-gate.md`, `subagent-dispatch.md`, `auto-transition.md`, `dirty-worktree.md`, `scripts.md`) for the heavy policies, and sections/tables are well organized. However, the long inline evidence-recording and check-policy passages (lines 186–192) are comparable content that got dedicated reference files elsewhere, so some separable content stays inline — anchor 4 ('good structure, minor organization gaps') rather than 5 or 3 (references are clearly signaled, not buried). | 4 / 5 |
Total | 16 / 20 Passed |